Adversarial Resilience Learning — Towards Systematic Vulnerability Analysis for Large and Complex Systems

Fischer, Lars and Memmen, Jan-Menno and Veith, Eric M. S. P. and Tröschel, Martin
ENERGY 2019, The Ninth International Conference on Smart Grids, Green Communications and IT Energy-aware Technologies
Abstract—Cyber-physical systems (CPSs) can be checked using numerous approaches, ranging from algorithmic model checking for a complete coverage of a finite-state system to extensive simulation, after which the system’s state is compared with defined invariants. However, modern CPSs are confronted with an increased amount of stochastic inputs, from volatile energy sources in power grids to broad user participation stemming from markets. The search space for a complete cover of a CPS becomes too large, while contracts cannot be formulated anymore considering the potentially erratic behavior of a user, or even in the face of a cyber attack. At the same time, the goal of resilience critical infrastructure cannot be eschewed, but the integration of user behavior and even non-checkable artificial intelligence algorithms is mandated, even required to meet, e.g., the goal to satisfy 80% of the gross power consumption from renewable energy sources by 2050. The concept of Adversarial Resilience Learning (ARL) formulates a new approach to CPS checking and resilient operation. It defines two agent classes, attacker and defender agents. The goal of the attacker is to de-stabilize the CPS, whereas the defender works to maintain a stable operational state. The quintessence of ARL lies in the attacker training the defender on a model of the CPS; as such, it is not a zero-sum game, but the learning of a resilient operation strategy for a CPS. This paper introduces the concept and the nomenclature of ARL, and, based on it, the description of experimental setups and results of a preliminary implementation of ARL in simulated power systems.
24 to 32